external-gitcode-ascend-ascendc-operator-precision-eval

Warn

Audited by Gen Agent Trust Hub on Apr 18, 2026

Risk Level: MEDIUMCOMMAND_EXECUTIONREMOTE_CODE_EXECUTIONPROMPT_INJECTION
Full Analysis
  • [COMMAND_EXECUTION]: Executes shell commands to configure environment variables and run testing tools. Evidence: source /usr/local/Ascend/ascend-toolkit/set_env.sh and python3 -m pytest in SKILL.md.
  • [REMOTE_CODE_EXECUTION]: Generates and executes Python scripts by populating templates with code snippets extracted from local project files. Evidence: Templates test_op_precision_template.py and run_precision_report_template.py contain placeholders like {{NPU_CALL}} and {{CPU_REF}} that are replaced with logic extracted from files like register.cpp or design.md.
  • [REMOTE_CODE_EXECUTION]: Performs dynamic loading of compiled C++ libraries from computed paths at runtime. Evidence: Template scripts search for and load .so files via torch.ops.load_library().
  • [PROMPT_INJECTION]: Subject to indirect prompt injection (Category 8) due to the ingestion of data from local project files into execution templates. Ingestion points: Reads implementation details and test cases from design.md, register.cpp, and <op_name>-test-cases.md. Boundary markers: None identified. Capability inventory: Shell execution, script generation, and library loading. Sanitization: No validation is performed on strings extracted from files before they are used as executable code.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Apr 18, 2026, 03:04 AM
Security Audit — agent-trust-hub — external-gitcode-ascend-ascendc-operator-precision-eval